Predicting Individuals at Risk of Diabetes Using Machine Learning.

Area of Science:

• * Medical Informatics
• * Public Health
• * Machine Learning in Healthcare

Background:

• * Diabetes mellitus (DM) poses a significant global health burden, necessitating early risk identification for effective prevention.
• * Machine learning (ML) offers a promising approach for timely diabetes detection and clinical decision support.
• * The study addresses the challenge of predicting diabetes risk within Mexico's primary healthcare system.

Purpose of the Study:

• * To develop and evaluate ML models for predicting diabetes risk using clinical and sociodemographic data.
• * To assess the performance of various supervised algorithms and ensemble techniques.
• * To address the critical issue of class imbalance in diabetes prediction.

Main Methods:

• * Analysis of data from 1,903 patients, including demographic, clinical variables, and health habits.
• * Implementation of supervised learning algorithms: Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), Multilayer Perceptron (MLP), and K-Nearest Neighbors (KNN).
• * Application of class balancing (Synthetic Minority Oversampling Technique - SMOTE) and ensemble methods (Extreme Gradient Boosting - XGBoost) to improve model performance.

Main Results:

• * Base ML models achieved high sensitivity (>90%) for identifying non-diabetic individuals (class 0) but low sensitivity (25-40%) for identifying at-risk individuals (class 1).
• * SMOTE and XGBoost significantly improved sensitivity for detecting diabetes risk to 79%, albeit with a decrease in specificity to 51.9%.
• * Class imbalance was confirmed as a critical challenge impacting the accurate prediction of diabetes risk.

Conclusions:

• * ML models demonstrate substantial potential for supporting diabetes risk identification, but class imbalance requires careful management.
• * Enhancing model sensitivity is crucial for promoting timely confirmatory testing and early intervention to prevent chronic complications.
• * Development of interpretable and clinically validated ML tools can improve patient-provider communication and adherence to preventive strategies.